@svgrid/grid 3.0.4 → 3.0.6

package/src/server-block-cache.ts

@@ -0,0 +1,690 @@
+/**
+ * Block cache for server-backed rows - the engine behind infinite scrolling.
+ *
+ * Paging asks for "page 3". Infinite scrolling asks a different question: the
+ * user is looking at rows 4,000 to 4,020 of a table with a million rows, so
+ * fetch the block that covers them, keep a bounded number of recently-seen
+ * blocks, and render something sensible for every row nobody has fetched yet.
+ * That is all this module does, in plain TypeScript with no Svelte and no DOM.
+ *
+ * The shape of the problem (and most of the vocabulary) is the same one every
+ * grid with a server row model solves, so the options here are deliberately
+ * recognisable: `blockSize`, `maxBlocksInCache`, `maxConcurrentRequests`,
+ * `blockLoadDebounceMs`, `initialRowCount`, `overflowRows`.
+ *
+ * Three things it takes seriously:
+ *
+ *   - **Unloaded rows still have to render.** `rows()` always returns a dense
+ *     array, with a shared sentinel object in every slot that has no data. The
+ *     grid recognises those (see {@link rowPlaceholderState}) and draws a
+ *     skeleton or a retry affordance instead of an empty row, so the scrollbar
+ *     never lies and scrolling never leaves a hole.
+ *   - **The row count may be unknown.** A backend that cannot cheaply count
+ *     says so (omit `rowCount`, or send `-1`) and the list grows by a block at
+ *     a time until a short block proves where the end is.
+ *   - **Requests are not free.** In-flight blocks are deduplicated, at most
+ *     `maxConcurrentRequests` are open at once, a fast scroll debounces rather
+ *     than firing a request per frame, and blocks abandoned by eviction or
+ *     purge are aborted.
+ *
+ * Enterprise's server row model runs one of these per group level; the free
+ * flat controller runs exactly one.
+ */
+
+/**
+ * The mark that says "this row is not data yet".
+ *
+ * A symbol, for two reasons. It cannot collide with a field in anyone's row
+ * data, which a string key like `__svPlaceholder` eventually would. And
+ * `Symbol.for` puts it in the global registry, so a placeholder still reads as
+ * one after it has crossed a module boundary - two copies of this file loaded
+ * through different specifiers would otherwise mint different symbols.
+ */
+const PLACEHOLDER = Symbol.for('svgrid.rowPlaceholder')
+
+/** Rows the cache has not loaded. One shared object per state, not per row. */
+const LOADING_ROW = Object.freeze({ [PLACEHOLDER]: 'loading' as const })
+const FAILED_ROW = Object.freeze({ [PLACEHOLDER]: 'failed' as const })
+
+/**
+ * Why a row is not real data, or `null` when it is.
+ *
+ * Deliberately NOT an identity check. The sentinels are shared objects, so
+ * `row === LOADING_ROW` looks like the obvious test - but rows routinely
+ * arrive here through something that wrapped them. Svelte is the everyday
+ * case: assigning the controller state into `$state` makes the row array
+ * deeply reactive, and every row read back out is a proxy of the original.
+ * Identity fails, every placeholder reads as real data, and the grid renders
+ * a screen of blank rows instead of skeletons. Reading the mark works through
+ * any wrapper that forwards property access, which is all of them.
+ */
+export function rowPlaceholderState(row: unknown): 'loading' | 'failed' | null {
+  if (row == null || typeof row !== 'object') return null
+  const mark = (row as Record<symbol, unknown>)[PLACEHOLDER]
+  return mark === 'loading' || mark === 'failed' ? mark : null
+}
+
+/**
+ * Mint a frozen placeholder the grid will recognise, carrying whatever
+ * other fields the caller wants on it. The row model uses this to give its
+ * placeholders a `kind`, so code reading display rows and code reading grid
+ * rows agree on what they are looking at.
+ */
+export function createRowPlaceholder<T extends object>(
+  state: 'loading' | 'failed',
+  fields: T,
+): Readonly<T & { readonly [PLACEHOLDER]: 'loading' | 'failed' }> {
+  return Object.freeze({ ...fields, [PLACEHOLDER]: state })
+}
+
+/** What a block fetch has to answer with. `rowCount` is optional: see {@link BlockCacheOptions}. */
+export type BlockFetchResult<TData> = {
+  rows: ReadonlyArray<TData>
+  /**
+   * Total rows after filtering. Omit it (or send `-1`) when the backend cannot
+   * count cheaply - the cache then discovers the end from the first short
+   * block, and `lastRowKnown()` stays false until it does.
+   */
+  rowCount?: number
+}
+
+/** What a block cache is built from: block size, caps, debounce and the fetch. */
+export type BlockCacheOptions<TData> = {
+  /**
+   * Rows per request. Bigger blocks mean fewer round trips and more wasted
+   * rows when the user scrolls past; 100 is the usual compromise.
+   */
+  blockSize?: number
+  /**
+   * Keep at most this many loaded blocks. Blocks outside the viewport are
+   * evicted least-recently-seen first, and scrolling back re-fetches them.
+   * Unlimited by default, which is right until the dataset is big enough that
+   * holding every visited block matters.
+   */
+  maxBlocksInCache?: number
+  /** Requests open at once. Default 2. */
+  maxConcurrentRequests?: number
+  /**
+   * Wait this long after the viewport settles before fetching. Non-zero values
+   * stop a drag of the scrollbar from requesting every block it flies past.
+   */
+  blockLoadDebounceMs?: number
+  /**
+   * Rows to claim before anything is loaded, so the grid has a scrollbar (and
+   * `scrollToRow` has somewhere to land) on the first paint. Default 1.
+   */
+  initialRowCount?: number
+  /**
+   * While the total is unknown, how many unloaded rows to keep past the last
+   * loaded one. Scrolling into them is what asks for the next block. Default 1.
+   */
+  overflowRows?: number
+  /** Fetch one block. Reject, or throw, to mark it failed. */
+  fetch: (
+    startRow: number,
+    endRow: number,
+    signal: AbortSignal,
+  ) => Promise<BlockFetchResult<TData>>
+  /** Called after any change to the rows, the count, or a block's state. */
+  onChange: (cache: BlockCacheState) => void
+  /**
+   * How to coalesce `onChange`. Defaults to a microtask, so a burst of block
+   * arrivals rebuilds the row array once. Pass `(fn) => fn()` in tests to make
+   * every change synchronous.
+   */
+  schedule?: (flush: () => void) => void
+}
+
+/** One block's place in the world, for diagnostics and for tests. */
+export type BlockState = {
+  blockIndex: number
+  startRow: number
+  endRow: number
+  status: 'loading' | 'loaded' | 'failed'
+  /** Monotonic counter of when the viewport last covered this block. */
+  lastTouched: number
+}
+
+/** One block as `getCacheState()` reports it. */
+export type BlockCacheState = {
+  /** Rows after filtering, or `null` while the backend has not said. */
+  rowCount: number | null
+  /** False while the end of the data is still being discovered. */
+  lastRowKnown: boolean
+  /** True while at least one block is in flight. */
+  loading: boolean
+  /** Blocks whose fetch rejected. Empty unless something went wrong. */
+  failedBlocks: number[]
+}
+
+/**
+ * A block cache: rows arrive in blocks as the viewport reaches them,
+ * placeholders stand in until then, and blocks far from the viewport are
+ * evicted past the cap.
+ */
+export type BlockCache<TData> = {
+  /**
+   * Tell the cache which rows are on screen. Fetches what is missing (after
+   * `blockLoadDebounceMs`) and marks those blocks as recently seen so eviction
+   * spares them. Safe to call on every scroll frame.
+   */
+  setViewport(startIndex: number, endIndex: number): void
+  /** The dense row array to hand the grid. Placeholders fill unloaded slots. */
+  rows(): ReadonlyArray<TData>
+  /** One row, without building the array. Returns a placeholder when unloaded. */
+  getRow(index: number): TData | typeof LOADING_ROW | typeof FAILED_ROW
+  rowCount(): number | null
+  lastRowKnown(): boolean
+  /** Re-fetch the blocks that failed. */
+  retryFailed(): void
+  /**
+   * Re-fetch the failed block holding `rowIndex`, and only that one; the
+   * Retry on a failed row. A row whose block has not failed is a no-op.
+   */
+  retryFailedAt(rowIndex: number): void
+  /**
+   * Re-fetch the loaded blocks in place, keeping the row count and scroll
+   * position. What you want after a mutation lands on the server.
+   */
+  refresh(): void
+  /** Drop everything and start over from the current viewport. */
+  purge(): void
+  /** Override the total, e.g. from a count endpoint that answered separately. */
+  setRowCount(count: number | null, known?: boolean): void
+  /**
+   * Write rows straight into the cache from `startRow`, as if a fetch had
+   * returned them, bypassing the datasource, the debounce and the
+   * concurrency cap. `startRow` must sit on a block boundary. An explicit
+   * `rowCount` settles the total; without one a short final slice does.
+   */
+  applyRows(startRow: number, rows: ReadonlyArray<TData>, rowCount?: number): void
+  /** Replace one loaded row. No-op when its block is not loaded. */
+  patch(index: number, row: TData): boolean
+  /**
+   * The index of the first LOADED row matching `predicate`, or -1. Rows
+   * that are not loaded are not visited - a transaction addressed by id
+   * can only touch what is in the cache.
+   */
+  findIndex(predicate: (row: TData, index: number) => boolean): number
+  /**
+   * Insert rows at an index and grow the count. Works within whichever
+   * run of loaded blocks contains `index`; loaded blocks AFTER that run
+   * are dropped, because their rows have shifted and will be re-fetched at
+   * their new offsets. An index in an unloaded region grows the count
+   * only (the rows exist; they arrive when scrolled to). Returns whether
+   * the rows were placed in the cache.
+   */
+  insert(index: number, rows: ReadonlyArray<TData>): boolean
+  /**
+   * Remove `count` rows at an index and shrink the count, with the same
+   * run semantics as `insert`. Returns how many rows were actually
+   * removed from the cache.
+   */
+  remove(index: number, count?: number): number
+  /** Every block the cache is holding, for `debug` output and tests. */
+  getCacheState(): BlockState[]
+  /** Abort what is in flight and stop emitting. Call on unmount. */
+  dispose(): void
+}
+
+type Block<TData> = {
+  status: 'loading' | 'loaded' | 'failed'
+  rows: TData[]
+  lastTouched: number
+  controller: AbortController | null
+}
+
+const defaultSchedule = (flush: () => void): void => {
+  void Promise.resolve().then(flush)
+}
+
+/**
+ * Build a block cache over a `fetch(startRow, endRow, signal)`. The free
+ * infinite row model and the Enterprise server-side row model both run on
+ * it, one per level.
+ */
+export function createBlockCache<TData>(options: BlockCacheOptions<TData>): BlockCache<TData> {
+  const blockSize = Math.max(1, options.blockSize ?? 100)
+  const maxBlocks = options.maxBlocksInCache ?? Infinity
+  const maxConcurrent = Math.max(1, options.maxConcurrentRequests ?? 2)
+  const debounceMs = Math.max(0, options.blockLoadDebounceMs ?? 0)
+  const initialRowCount = Math.max(0, options.initialRowCount ?? 1)
+  const overflowRows = Math.max(0, options.overflowRows ?? 1)
+  const schedule = options.schedule ?? defaultSchedule
+
+  const blocks = new Map<number, Block<TData>>()
+  /** Block indices waiting for a slot, in the order the viewport asked for them. */
+  const queue: number[] = []
+  let inFlight = 0
+  let touchClock = 0
+  let count: number | null = null
+  let countKnown = false
+  let disposed = false
+
+  let viewStart = 0
+  let viewEnd = 0
+  let debounceTimer: ReturnType<typeof setTimeout> | null = null
+
+  let emitScheduled = false
+  let rowsCache: TData[] | null = null
+
+  const blockOf = (rowIndex: number): number => Math.floor(rowIndex / blockSize)
+  const startOf = (blockIndex: number): number => blockIndex * blockSize
+
+  /**
+   * How many rows to claim we have.
+   *
+   * With a known count that is simply the count. Without one it is the last row
+   * we have touched plus `overflowRows`, so there is always somewhere to scroll
+   * that will ask for the next block - this is what makes the list grow as you
+   * go rather than ending at the first unloaded row.
+   */
+  function currentLength(): number {
+    if (countKnown && count != null) return count
+    let highestEnd = 0
+    let anySettled = false
+    for (const b of blocks.values()) if (b.status !== 'loading') anySettled = true
+    for (const [index, block] of blocks) {
+      // A block in flight extends the list only once some block has settled.
+      // Before that the list is `initialRowCount` long: the few skeleton
+      // rows a level shows while it waits, not a hundred of them. The first
+      // block failing says nothing about how long the list is either, so it
+      // keeps those few rows (its Retry sits on them); a failed block further
+      // in was scrolled to, so its slots stay and the scrollbar holds still.
+      const end = block.status === 'loaded'
+        ? startOf(index) + block.rows.length
+        : block.status === 'failed'
+          ? index === 0 ? 0 : startOf(index) + blockSize
+          : anySettled
+            ? startOf(index) + blockSize
+            : 0
+      if (end > highestEnd) highestEnd = end
+    }
+    if (!anySettled) return initialRowCount
+    return Math.max(initialRowCount, highestEnd + overflowRows)
+  }
+
+  function state(): BlockCacheState {
+    const failed: number[] = []
+    for (const [index, block] of blocks) if (block.status === 'failed') failed.push(index)
+    return {
+      rowCount: countKnown ? count : null,
+      lastRowKnown: countKnown,
+      loading: inFlight > 0,
+      failedBlocks: failed.sort((a, b) => a - b),
+    }
+  }
+
+  /**
+   * Something changed: forget the row array NOW, so a synchronous read after
+   * a mutation sees the new state, and notify on the next tick, so a burst of
+   * mutations notifies once. Every mutating path ends here.
+   */
+  function emit(): void {
+    rowsCache = null // rebuilt lazily, on the next read
+    if (disposed || emitScheduled) return
+    emitScheduled = true
+    schedule(() => {
+      emitScheduled = false
+      if (disposed) return
+      options.onChange(state())
+    })
+  }
+
+  function ensureBlock(blockIndex: number): void {
+    if (disposed) return
+    const existing = blocks.get(blockIndex)
+    // Loading or loaded blocks are left alone; a failed one only retries when
+    // asked, so a broken backend is not hammered once per scroll frame.
+    if (existing) return
+    if (queue.includes(blockIndex)) return
+    queue.push(blockIndex)
+    pump()
+  }
+
+  function pump(): void {
+    while (!disposed && inFlight < maxConcurrent && queue.length > 0) {
+      const blockIndex = queue.shift()!
+      if (blocks.has(blockIndex)) continue
+      void load(blockIndex)
+    }
+  }
+
+  async function load(blockIndex: number): Promise<void> {
+    const controller = new AbortController()
+    const block: Block<TData> = {
+      status: 'loading',
+      rows: [],
+      lastTouched: (touchClock += 1),
+      controller,
+    }
+    blocks.set(blockIndex, block)
+    inFlight += 1
+    emit()
+
+    const startRow = startOf(blockIndex)
+    const endRow = startRow + blockSize
+    try {
+      const result = await options.fetch(startRow, endRow, controller.signal)
+      if (disposed || controller.signal.aborted || blocks.get(blockIndex) !== block) return
+      block.rows = [...result.rows]
+      block.status = 'loaded'
+      block.controller = null
+      applyCountFrom(blockIndex, block.rows.length, result.rowCount)
+      evictIfNeeded()
+    } catch {
+      if (disposed || controller.signal.aborted || blocks.get(blockIndex) !== block) return
+      block.status = 'failed'
+      block.rows = []
+      block.controller = null
+    } finally {
+      if (!disposed && !controller.signal.aborted) {
+        inFlight -= 1
+        emit()
+        pump()
+      }
+    }
+  }
+
+  /**
+   * Learn the total from a block that just landed.
+   *
+   * An explicit non-negative `rowCount` settles it. Otherwise a SHORT block is
+   * the proof that we hit the end: fewer rows came back than were asked for, so
+   * the data stops where they stop.
+   */
+  function applyCountFrom(blockIndex: number, received: number, reported: number | undefined): void {
+    if (typeof reported === 'number' && reported >= 0) {
+      count = reported
+      countKnown = true
+      return
+    }
+    if (received < blockSize) {
+      count = startOf(blockIndex) + received
+      countKnown = true
+    }
+  }
+
+  /**
+   * Evict loaded blocks over the limit, least-recently-seen first, never one
+   * the viewport is currently on - dropping a visible block would swap real
+   * rows for skeletons under the user's eyes and immediately re-fetch them.
+   */
+  function evictIfNeeded(): void {
+    if (!Number.isFinite(maxBlocks)) return
+    const firstVisible = blockOf(viewStart)
+    const lastVisible = blockOf(Math.max(viewStart, viewEnd))
+    const evictable = [...blocks.entries()]
+      .filter(([index, b]) => b.status === 'loaded' && (index < firstVisible || index > lastVisible))
+      .sort((a, b) => a[1].lastTouched - b[1].lastTouched)
+
+    let loaded = 0
+    for (const b of blocks.values()) if (b.status === 'loaded') loaded += 1
+
+    for (const [index] of evictable) {
+      if (loaded <= maxBlocks) break
+      blocks.delete(index)
+      loaded -= 1
+    }
+  }
+
+  function fetchViewport(): void {
+    if (disposed) return
+    const first = blockOf(viewStart)
+    const last = blockOf(Math.max(viewStart, viewEnd))
+    for (let i = first; i <= last; i += 1) {
+      const block = blocks.get(i)
+      if (block) block.lastTouched = touchClock += 1
+      else ensureBlock(i)
+    }
+  }
+
+  function abortAll(): void {
+    for (const block of blocks.values()) block.controller?.abort()
+    inFlight = 0
+    queue.length = 0
+  }
+
+  return {
+    setViewport(startIndex, endIndex) {
+      if (disposed) return
+      viewStart = Math.max(0, startIndex)
+      viewEnd = Math.max(viewStart, endIndex)
+      if (debounceMs === 0) {
+        fetchViewport()
+        return
+      }
+      if (debounceTimer) clearTimeout(debounceTimer)
+      debounceTimer = setTimeout(() => {
+        debounceTimer = null
+        fetchViewport()
+      }, debounceMs)
+    },
+
+    rows() {
+      if (rowsCache) return rowsCache
+      const length = currentLength()
+      const out = new Array<TData>(length)
+      // Fill from the blocks we have, then pad the rest with placeholders. One
+      // pass per block beats one lookup per row: at a million rows the
+      // difference is the frame budget.
+      out.fill(LOADING_ROW as unknown as TData)
+      for (const [index, block] of blocks) {
+        const start = startOf(index)
+        if (start >= length) continue
+        if (block.status === 'failed') {
+          const end = Math.min(start + blockSize, length)
+          for (let i = start; i < end; i += 1) out[i] = FAILED_ROW as unknown as TData
+          continue
+        }
+        for (let i = 0; i < block.rows.length && start + i < length; i += 1) {
+          out[start + i] = block.rows[i]!
+        }
+      }
+      rowsCache = out
+      return out
+    },
+
+    getRow(index) {
+      const block = blocks.get(blockOf(index))
+      if (!block) return LOADING_ROW
+      if (block.status === 'failed') return FAILED_ROW
+      const row = block.rows[index - startOf(blockOf(index))]
+      return row ?? LOADING_ROW
+    },
+
+    rowCount: () => (countKnown ? count : null),
+    lastRowKnown: () => countKnown,
+
+    retryFailed() {
+      let any = false
+      for (const [index, block] of [...blocks]) {
+        if (block.status !== 'failed') continue
+        blocks.delete(index)
+        ensureBlock(index)
+        any = true
+      }
+      if (any) emit()
+    },
+
+    retryFailedAt(rowIndex) {
+      const index = blockOf(Math.max(0, rowIndex))
+      if (blocks.get(index)?.status !== 'failed') return
+      blocks.delete(index)
+      ensureBlock(index)
+      emit()
+    },
+
+    refresh() {
+      // Keep the count and the scroll position: drop the DATA, re-request the
+      // same blocks. The user sees skeletons where they were, not a jump home.
+      for (const [index, block] of [...blocks]) {
+        block.controller?.abort()
+        blocks.delete(index)
+        ensureBlock(index)
+      }
+      emit()
+    },
+
+    purge() {
+      abortAll()
+      blocks.clear()
+      count = null
+      countKnown = false
+      rowsCache = null
+      fetchViewport()
+      emit()
+    },
+
+    setRowCount(next, known = next != null) {
+      count = next
+      countKnown = known && next != null
+      emit()
+    },
+
+    applyRows(startRow, rows, rowCount) {
+      if (startRow % blockSize !== 0) {
+        throw new Error(
+          `createBlockCache.applyRows: startRow ${startRow} is not a multiple of blockSize ${blockSize}`,
+        )
+      }
+      const firstBlock = blockOf(startRow)
+      const blockCount = Math.ceil(rows.length / blockSize)
+      for (let i = 0; i < blockCount; i += 1) {
+        const index = firstBlock + i
+        blocks.get(index)?.controller?.abort() // ours now, whatever was in flight
+        blocks.set(index, {
+          status: 'loaded',
+          rows: rows.slice(i * blockSize, (i + 1) * blockSize),
+          lastTouched: (touchClock += 1),
+          controller: null,
+        })
+      }
+      const lastBlock = blocks.get(firstBlock + blockCount - 1)
+      if (lastBlock) applyCountFrom(firstBlock + blockCount - 1, lastBlock.rows.length, rowCount)
+      else if (typeof rowCount === 'number' && rowCount >= 0) {
+        count = rowCount
+        countKnown = true
+      }
+      emit()
+    },
+
+    patch(index, row) {
+      const block = blocks.get(blockOf(index))
+      if (!block || block.status !== 'loaded') return false
+      const offset = index - startOf(blockOf(index))
+      if (offset < 0 || offset >= block.rows.length) return false
+      block.rows[offset] = row
+      emit()
+      return true
+    },
+
+    findIndex(predicate) {
+      for (const [blockIndex, block] of blocks) {
+        if (block.status !== 'loaded') continue
+        const start = startOf(blockIndex)
+        for (let i = 0; i < block.rows.length; i += 1) {
+          if (predicate(block.rows[i]!, start + i)) return start + i
+        }
+      }
+      return -1
+    },
+
+    insert(index, rows) {
+      if (rows.length === 0) return false
+      const clamped = Math.max(0, Math.min(index, currentLength()))
+      const run = loadedRunAt(clamped)
+      if (run) {
+        // Rows shift across block boundaries, so re-slice the run from its
+        // first block; everything loaded past it is dropped (offsets moved).
+        run.flat.splice(clamped - startOf(run.startBlock), 0, ...rows)
+        reseat(run.startBlock, run.flat)
+      } else {
+        dropLoadedFrom(blockOf(clamped))
+      }
+      if (countKnown && count != null) count += rows.length
+      emit()
+      return run != null
+    },
+
+    remove(index, removeCount = 1) {
+      if (removeCount <= 0) return 0
+      const run = loadedRunAt(index)
+      if (!run) return 0
+      const removed = run.flat.splice(index - startOf(run.startBlock), removeCount).length
+      reseat(run.startBlock, run.flat)
+      if (countKnown && count != null) count = Math.max(0, count - removed)
+      emit()
+      return removed
+    },
+
+    getCacheState() {
+      return [...blocks.entries()]
+        .map(([blockIndex, block]) => ({
+          blockIndex,
+          startRow: startOf(blockIndex),
+          endRow: startOf(blockIndex) + blockSize,
+          status: block.status,
+          lastTouched: block.lastTouched,
+        }))
+        .sort((a, b) => a.blockIndex - b.blockIndex)
+    },
+
+    dispose() {
+      disposed = true
+      if (debounceTimer) clearTimeout(debounceTimer)
+      debounceTimer = null
+      abortAll()
+      blocks.clear()
+    },
+  }
+
+  /**
+   * The run of consecutive loaded blocks containing `index`, flattened, with
+   * the block it starts at. Null when the block at `index` is not loaded -
+   * an index-shifting edit there would have nothing to shift.
+   */
+  function loadedRunAt(index: number): { startBlock: number; flat: TData[] } | null {
+    const at = blockOf(index)
+    if (blocks.get(at)?.status !== 'loaded') return null
+    let startBlock = at
+    while (blocks.get(startBlock - 1)?.status === 'loaded') startBlock -= 1
+    const flat: TData[] = []
+    for (let i = startBlock; ; i += 1) {
+      const block = blocks.get(i)
+      if (!block || block.status !== 'loaded') break
+      flat.push(...block.rows)
+      if (block.rows.length < blockSize) break // the last block, by definition
+    }
+    return { startBlock, flat }
+  }
+
+  /**
+   * Lay a flat run back out into blocks from `startBlock`, and drop every
+   * block after the run: their rows have shifted, and re-fetching them at
+   * the new offsets is the honest outcome. Blocks before it are untouched.
+   */
+  function reseat(startBlock: number, flat: TData[]): void {
+    const blockCount = Math.ceil(flat.length / blockSize)
+    dropLoadedFrom(startBlock)
+    for (let i = 0; i < blockCount; i += 1) {
+      blocks.set(startBlock + i, {
+        status: 'loaded',
+        rows: flat.slice(i * blockSize, (i + 1) * blockSize),
+        lastTouched: (touchClock += 1),
+        controller: null,
+      })
+    }
+    rowsCache = null
+  }
+
+  /** Forget every block from `fromBlock` on, aborting any in flight. */
+  function dropLoadedFrom(fromBlock: number): void {
+    for (const [index, block] of [...blocks]) {
+      if (index < fromBlock) continue
+      block.controller?.abort()
+      blocks.delete(index)
+    }
+    rowsCache = null
+  }
+}